基于改进Cascade R-CNN算法的船舶目标检测方法  

Boat object detection method based on improved Cascade R-CNN algorithm

作　　者：杨镇宇 石刘[1] YANG Zhen-yu;SHI Liu(China Ship Research and Development Academy,Beijing 100192,China)

机构地区：[1]中国舰船研究院,北京100192

出　　处：《舰船科学技术》2024年第6期144-149,共6页Ship Science and Technology

摘　　要：为了解决实际场景下船舶目标检测精度低的问题,本文在Cascade R-CNN算法的基础上,提出一种船舶目标检测方法 Boat R-CNN。Boat R-CNN使用带自注意力机制的Swin-Transformer Tiny网络提取图像特征,使用Soft-NMS非极大值抑制方法提升候选框过滤精度,使用Smooth_L1损失函数加速模型收敛并减少梯度爆炸情况,使用CIOU边界框回归损失提高候选框回归质量,并针对船舶目标的形状特征优化锚框的长宽比,提高锚框的生成质量。实验结果表明,Boat R-CNN算法的精度相比原版Cascade R-CNN算法提高了21.8%,相比主流Faster R-CNN算法提高了30.3%,有效提升了实际场景下的船舶目标检测精度。To address the issue of low accuracy in boat object detection in real-world scenarios,this paper improves upon the Cascade R-CNN algorithm and proposes a boat object detection method called Boat R-CNN.Boat R-CNN utilizes the Swin-Transformer Tiny network with a self-attention mechanism to extract image features,employs Soft-NMS for nonmaximum suppression to enhance the filtering precision of candidate bounding boxes,uses the Smooth_L1 loss function to accelerate model convergence and reduce gradient explosion,and utilizes CIOU bounding box regression loss to improve the quality of candidate box regression.Furthermore,the aspect ratio of anchor boxes is optimized for the shape characteristics of boat objects,improving the quality of anchor box generation.Experimental results have shown that the Boat R-CNN algorithm has increased accuracy by 21.8% compared to the original Cascade R-CNN algorithm and 30.3% compared to the mainstream Faster R-CNN algorithm.Boat R-CNN effectively improves the accuracy of boat object detection in real-world scenarios.

关 键 词：船舶 目标检测 深度学习 Cascade R-CNN Swin Transformer 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术]